OpenAI Says Its Model Solved 100+ Open Math Problems, Forms Advisory Group

OpenAI's math story just got much bigger, and it comes with new adult supervision. On September 21, 2026, OpenAI published Advisory Group on Mathematics and Artificial Intelligence, confirming that an internal model it began training on August 28 has, beyond the Navier-Stokes Millennium Prize problem announced two weeks earlier, now resolved more than 100 long-standing open problems across most areas of mathematics.

That is not a benchmark score. It is OpenAI's own characterization of internal results that have not been independently verified, named, or published in detail. The same post announces something more concrete and checkable: an independent advisory group hosted at the Institute for Advanced Study, with nine mathematicians including Fields Medalist-adjacent heavyweights Edward Witten and Timothy Gowers, formed in direct response to the Fields Medalists' "severe misalignment" criticism from ten days earlier.

TL;DR — what people are actually asking

What OpenAI's post actually says

The post's own framing, worth quoting directly:

On August 28, we began training a new internal model. In addition to resolving the Navier-Stokes Millennium Prize problem, this model has now resolved more than 100 long-standing open problems across most areas of mathematics. The pace of its progress in mathematics has surprised the mathematicians within OpenAI.

Read carefully, that sentence covers a single continuous effort, not two separate models. The Navier-Stokes result announced September 8 — the one that triggered Tristan Buckmaster's public credit dispute and Terence Tao and 24 other Fields Medalists' formal objection — came from the same training run that OpenAI now says has gone on to clear 100+ additional problems. OpenAI frames its own internal mathematicians as surprised by the pace, which is either a genuine research surprise or a carefully chosen way to preempt the "why didn't you disclose this sooner" question the community had just raised about the last result.

Crucially, OpenAI explicitly names the mechanism for its own caution here: the open letter titled "A Severe Misalignment of AI in Mathematics" is cited by name in the post as the trigger for "internal discussions on the best way to inform the community of the rapid progress to prepare and adapt the field." That is an unusually direct acknowledgment that outside criticism changed how a frontier lab is communicating a result — not just cosmetically, but structurally, via a new standing body.

The advisory group: what it can and can't do

The Advisory Group on Mathematics and Artificial Intelligence is hosted at the Institute for Advanced Study — the Princeton institution where Einstein and Gödel once worked, and a genuinely prestigious host for a corporate advisory body. Its nine initial members:

Gowers is himself a Fields Medalist and a co-signer of the September 11 declaration OpenAI is responding to — his presence on the advisory board is the clearest single signal that this isn't a hand-picked friendly panel. Witten is one of the most decorated physicists alive and a Fields Medal recipient in his own right for work bridging physics and mathematics.

OpenAI's post is specific about the group's scope and its limits:

• What it does: advise on review and communication of emerging results, help assess significance, advise on coordinating dissemination, advise on academic and professional research standards, and advise on how OpenAI's tools can support mathematical research and learning.
• What it explicitly does not do: "the group will not be responsible for advising us on how to pace our internal progress on mathematics." OpenAI keeps full control over what it builds and when — the group's remit is communication and standards, not the research roadmap.
• Independence claims: the group operates independently, can offer advice OpenAI hasn't requested, can comment publicly on OpenAI's impact on mathematics, and can make its advice public. Members are unpaid and can change the group's own membership.

That last set of claims is the part worth watching over time rather than taking at face value today. An advisory group that can genuinely criticize its host publicly, with no pay and no OpenAI control over pacing, is structurally different from a typical corporate "ethics board." Whether it behaves that way in practice — publishing dissent, not just being cited as evidence of good governance — is the test the next few months will run.

Why this follows directly from the Fields Medalists' letter

The timeline matters here:

1. September 8 — OpenAI announces the Navier-Stokes result, a 10,000-agent swarm effort run over 88 hours.
2. Within a day — Tristan Buckmaster publicly disputes OpenAI's independence claim, alleging the effort was triggered by his and Anthropic researcher Levent Alpöge's private research.
3. September 11 — 25 Fields Medalists, including Terence Tao, sign "A Severe Misalignment of AI in Mathematics", arguing AI labs are treating famous open problems as PR benchmarks at the expense of the field's actual process.
4. September 16 — Scott Aaronson relays an unverified rumor that labs, "burned by the hostile response," may now be sitting on solved problems rather than announcing them.
5. September 21 — OpenAI confirms 100+ additional resolved problems and forms the IAS-hosted advisory group, citing the September 11 letter directly.

Read against that sequence, this post does two things at once that could look contradictory: it discloses more math results than anyone expected (cutting against the "labs are hoarding" theory Aaronson's rumor floated five days earlier), while simultaneously building the institutional apparatus the Fields Medalists' letter demanded. If Aaronson's rumor had any truth to it, this looks like OpenAI choosing disclosure-with-oversight over silence — though it's equally consistent with a lab timing a large reveal to land alongside a credibility-repair announcement rather than on its own.

What's confirmed and what isn't

Confirmed, directly from OpenAI's own post:

• The advisory group exists, is hosted at IAS, and has nine named members.
• OpenAI states publicly that the group will not control its pacing decisions.
• OpenAI says an internal model trained since August 28 resolved Navier-Stokes and claims 100+ other problems.

Not confirmed, not disclosed in this post:

• Which 100+ problems were resolved, or any proofs, papers, or Lean formalizations for them.
• Any independent mathematician's verification of the 100+ figure — the advisory group's formation is announced in the same post as the claim, not after a review of it.
• Whether the Buckmaster/Alpöge credit dispute over the original Navier-Stokes result has been addressed, apologized for, or resolved in any way — the post doesn't mention it.
• A timeline for when specific results will be published for outside scrutiny.

That gap between "we did this" and "here's the proof, go check it" is exactly the pattern the Fields Medalists' letter objected to in the first place. Forming an advisory group to fix a communication problem, in a post that repeats the same disclosure pattern the group exists to fix, is worth sitting with rather than accepting at face value.

What this means if you follow AI-and-math claims

For anyone tracking frontier-lab math claims rather than doing frontier math research directly, three things from this episode are actionable:

• A named, credentialed advisory board is a stronger signal than a launch-post claim, but it is not verification. Gowers signing both the critical letter and joining the board he's now part of is a genuinely unusual credibility move — track whether the group publishes anything critical in the coming months, not just whether OpenAI cites its existence.
• "100+ problems" with zero named examples is a number to discount until specifics land. The same caution that applied to the original Navier-Stokes claim applies here, scaled up: a bigger unverified number is not more credible than a smaller one, just louder.
• Watch whether Aaronson's hoarding rumor gets addressed directly. OpenAI's post reads as circumstantial evidence against the hoarding claim, but it doesn't name or respond to Aaronson's rumor at all — a real answer to that specific rumor would be a stronger signal than a large, separate disclosure landing in the same week.

Related on explainx.ai

Update — September 22, 2026: OpenAI has formed an independent advisory group and confirmed 100+ additional resolved problems — see this post for the full response to the criticism below.

• OpenAI's Navier-Stokes Proof Is Now a Credit and Data Dispute — the original result and the still-unresolved Buckmaster/Alpöge allegation
• 25 Fields Medalists Accused AI Labs of "Severe Misalignment" in Math — the September 11 letter OpenAI is directly responding to
• Are AI Labs Now Hoarding Solved Math Problems to Avoid Backlash? — the September 16 rumor this announcement partially cuts against
• The 7 Millennium Prize Problems: What AI Has Actually Solved — the honest scorecard this post's 100+ claim should be measured against
• OpenAI Used 10,000 AI Agents to Propose a Solution to a 90-Year Math Problem — a related large-scale agent-swarm math effort from the same period
• Did Claude Solve Navier-Stokes? The Millennium Prize Rumor, Fact-Checked — the earlier misattributed rumor this saga traces back to
• The Real Story in OpenAI's Navier-Stokes Proof: Formal Verification Got 10,000x Cheaper
• Will AI Replace Mathematicians? The IEEE "Big Mathematics" Debate

Primary source: OpenAI, Advisory Group on Mathematics and Artificial Intelligence (September 21, 2026)

This post reflects OpenAI's own September 21, 2026 announcement. The 100+ resolved-problems claim is OpenAI's self-report, not independently verified — no problem list or proofs accompanied this announcement. The advisory group's independence claims (unpaid membership, no OpenAI control over pacing, ability to publish unsolicited criticism) are as stated by OpenAI and have not yet been tested by any public action from the group. Follow @explainx_ai for updates.